Reviews: GANs Trained by a Two Time-Scale Update Rule Converge to a Local Nash Equilibrium

Generative adversarial networks (GANs) are turning out to be a very important advance in machine learning. Algorithms for training GANs have difficulties with convergence. The paper proposes a two time-scale update rule (TTUR) which is shown (proven) to converge under certain assumptions. Specifically, it shows that GAN Adam updates with TTUR can be expressed as ordinary differential equations, and therefore can be proved to converge using a similar approach as in Borkar' 1997 work. The recommendation is to use two different update rules for generator and discriminator, with the latter being faster, in order to have convergence guarantees.